Computer Vision Applications and their Ethical Risks in the Global SouthCharles-Olivier Dufresne-Camaro* Fanny Chevalier† Syed Ishtiaque Ahmed‡

∗†‡ Department of Computer Science, † Department of Statistical Sciences, University of Toronto

ABSTRACT

We present a study of recent advances in computer vision (CV) re-search for the Global South to identify the main uses of modern CVand its most significant ethical risks in the region. We review 55research papers and analyze them along three principal dimensions:where the technology was designed, the needs addressed by thetechnology, and the potential ethical risks arising following deploy-ment. Results suggest: 1) CV is most used in policy planning andsurveillance applications, 2) privacy violations is the most likelyand most severe risk to arise from modern CV systems designed forthe Global South, and 3) researchers from the Global North differfrom researchers from the Global South in their uses of CV to solveproblems in the Global South. Results of our risk analysis also differfrom previous work on CV risk perception in the West, suggestinglocality to be a critical component of each risk’s importance.

Index Terms: General and reference—Document types—Surveysand overviews; Computing methodologies—Artificial intelligence—Computer vision—Computer vision tasks; Social and profes-sional topics—User characteristics; Social and professional topics—Computing / technology policy;

1 INTRODUCTION

In recent years, computer vision (CV) systems have become in-creasingly ubiquitous in many parts of the world, with applica-tions ranging from assisted photography on smartphones, to drone-monitored agriculture. Current popular problems in the field includeautonomous driving, affective computing and general activity recog-nition; each application has the potential to significantly impact theway we live in society.

Computer vision has also been leveraged for more controversialapplications, such as sensitive attributes predictors for sexual orien-tation [80] and body weight [42], DeepFakes [78,81] and automatedsurveillance systems [53]. In particular, this last application hasbeen shown to have the potential to severely restrict people’s privacyand freedom, and compromise security. As one example, we notethe case of China and their current use of surveillance systems toidentify, track and target Uighurs [53], a minority group primarilyof Muslim faith. Similar systems with varying functionalities havealso been exported to countries such as Ecuador [54].

While advances in CV have led to considerable technologicalimprovements, they also raise a great deal of ethical issues thatcannot be disregarded. We stress that it is critical that the potentialsocietal implications are studied as new technologies are developed.We recognize that this is a particularly difficult endeavour, as oneoften has very little control over how technology is used downstream.Moreover, recent focus on designing for global markets makes it nearimpossible to consider all local uses and adaptations: a contribution

*e-mail: camaro@cs.toronto.edu†e-mail: fanny@cs.toronto.edu‡e-mail: ishtiaque@cs.toronto.edu

making a positive impact in the Global North may turn out to be aharmful, restrictive one in the Global South 1.

Several groups have come forward in the last decade to raiseethical concerns over certain smart systems — CV and general-AIsystems alike — with reports of discrimination [14, 15, 44, 48] andprivacy violations [26, 50, 58, 62]. One notable example comes fromJoy Buolamwini, who exposed discriminatory issues in several faceanalysis systems [15]. However, very few have explored the morefocused general problem of ethics in CV. We believe this to be animportant problem as restricting the scope to CV allows for moredomain-specific actionable solutions to emerge, while keeping sightof the big-picture problems in the field.

To our knowledge, only two attempts to study this problem havebeen undertaken in modern CV research [45, 74]. While insight-ful, these works paint ethical issues in CV with a coarse brush,discarding subtleties of how these risks may vary across differentlocal contexts. Furthermore, the surveyed literature in both workfocuses extensively on CV in the West and major tech hubs, raisingconcerns over the generalization of their findings in other distinctcommunities. Considering people in the Global South are importantusers of technology, most notably smartphones, we believe it is ofutmost importance to extend this study of CV-driven risks to theircommunities.

In this work, we conduct a survey of CV research literature withapplications in the Global South to gain a better understanding ofCV uses and the underlying problems in this area. While studyingresearch literature may provide a weaker understanding of the trueproblem space compared to studying existing systems in the field,we believe our method forms a strong preliminary step in studyingthis important problem. In doing so, we aim to provide furtherguidance to designers in building safe context-appropriate solutionswith CV components.

To guide our search in identifying potential risks, we use Skirpanand Yeh’s moral compass [74] (Sect. 2). Through our analysis(Sect. 3 -5), we aim to answer the following research questions:

RQ1. What are the main applications of modern computer visionwhen used in the Global South?

RQ2. What are the most significant computer vision-driven ethicalrisks in a context of use in the Global South?

RQ3. How do research interests of Global North researchers dif-fer from those of Global South researchers in the context ofcomputer vision research for the Global South?

RQ4. How do the risks associated with these technologies when usedin the Global South differ from those in the Global North?

It is our belief that while certain CV technologies and algorithmsare equally available to people in the North and South, the resultinguses, implications, and risks will differ. Our contributions are four-fold: 1) we present an overview of research where CV technologyis deployed or studied in a geographical context within the Global

1We use the United Nations’ M49 Standard [77] to define the GlobalNorth (developed countries) and the Global South (developing countries).Graphics Interface Conference 2020

28-29 May Copyright held by authors. Permission granted to CHCCS/SCDHM to publish in print and digital form, and ACM to publish electronically.

South, and the needs it aims to address; 2) we identify the princi-pal ethical risks at play in recent CV research in this context; 3)we show that research interests in CV applications for the GlobalSouth differ between Global South and Global North researchers;and 4) we show that the importance of ethical risks in CV systemsdiffers across contexts. Finally, we also reflect on our findings andpropose design considerations to minimize risks in CV systems forthe Global South.

2 FRAMEWORK

We are aware of only two studies aiming at identifying ethical risksin computer vision. The first, by Lauronen [45], surveys recent CVliterature and briefly identifies six themes of ethical issues in CV:espionage, identity theft, malicious attacks, copyright infringement,discrimination, and misinformation.

The second, by Skirpan and Yeh [74], takes the form of a moralcompass (framework) aimed at guiding CV researchers in protectingthe public. In this framework, CV risks fall into five categories:privacy violations, discrimination, security breaches, spoofing andadversarial inputs, and psychological harms. Using those categories,the authors propose 22 risk scenarios based on recent research andreal-life events. In addition, they present a likelihood of occurrencerisk evaluation along with a preliminary risk perception evaluationbased on uncertainty (i.e. how much is known about it) and severity(i.e. how impactful it is).

To guide our search in evaluating CV in a Global South con-text, we propose using Skirpan and Yeh’s framework based on tworeasons: 1) we find the set of risk categories to generalize betterin non-Western countries, and 2) it allows us to directly comparethe results of our risk analysis in the Global South to the resultsof Skirpan and Yeh’s risk analysis in the West [74]. Nevertheless,we emphasize that the risk categories are by no mean exhaustive ofevery current risk related to modern CV systems. Indeed, the fieldis still evolving at a considerable rate and new risks may not all fitwell into any of those categories, e.g. DeepFakes.

To illustrate the scope of each risk category, we present belowone risk scenario extracted from the framework [74] along witha brief description of the risk. Note that we make no distinctionbetween deliberate and accidental presences of a risk, e.g. unplanneddiscrimination caused by the use of biased data to train a CV system.

Privacy violations

“Health insurance premium is increased due to infer-ences from online photos.” — Scenario #1 [74]

Privacy violations are defined as having one’s private informationcollected without the person’s consent by a third party via the use ofa CV system. This includes data such as body measurements, sexualorientation, income, relationships, travels and activities.

Discrimination

“Xenophobia leads police forces to track and target for-eigners.” — Scenario #9 [74]

Discrimination occurs when someone is treated unjustly by a CVsystem due to certain sensitive characteristics that are irrelevant tothe situation. Such characteristics include religious beliefs, race,gender and sexual orientation.

Security breaches

“Security guard sells footage of public official typing ina password to allow for a classified information leak.”— Scenario #5 [74]

Skirpan and Yeh broadly define security breaches as a maliciousactor(s) exploiting vulnerabilities in CV systems to disable them,steal private information or employ them in cyberattacks.

Spoofing and adversarial inputs

“Automated public transportation that uses visual verifi-cation systems is attacked causing a crash.”— Scenario #8 [74]

In this context, the objective of spoofing and adversarial inputs isto push a CV system to react confidently in an erroneous and harm-ful manner. Examples include misclassifying people or situations,e.g. recognizing a robber as a bank employee, or misclassifyingfraudulent documents as authentic.

Psychological harms

“People will not attend protests they agree with due tofear of recognition by cameras and subsequent punish-ment.” — Scenario #2 [74]

At a high level, any negative change in how people live based ontheir interactions with CV systems is considered a psychologicalharm. Unlike the other categories, this risk is more passive andresults from longer-term interactions with CV systems.

We stress that the risk scenarios and their analysis were aimed tobe general and reflect ”normal” life in the Global North. Therefore,they likely do not generalize well to a myriad of areas in the GlobalSouth. For example, scenario #1 used to illustrate privacy violationstakes for granted that access to health insurance is available andthat people have some sort of online presence. It also assumesthat a health insurance company has access to enough data on thepopulation to do any kind of reliable inference.

For this reason, we only use the risk categories to guide ouranalysis. We believe those generalize well in all areas with CVsystems, even though they may not cover the full risk space andwhat constitutes sensitive or private information may vary acrosscommunities [2, 6].

3 METHODOLOGY

In the context of this work, we consider a prior work (system) to berelevant to our study if it identifies or addresses specific needs ofGlobal South communities through the use of CV techniques. Thiscriterion can be satisfied in multiple ways: direct deployment of asystem in the area, use of visual data related to the area, descriptionof local problems related to CV research, and so forth. Worksexploring specific problems in more general settings are excluded,e.g. discrimination at large in CV-driven surveillance systems. Weexclude from our study systems designed for China, given theircurrent economical situation.

To find relevant research papers, we conducted a search on fourdigital libraries: ACM Digital Library, IEEE Xplore Digital Library,ScienceDirect and JSTOR. We have chosen these four platforms tomaximize our search coverage on both applied research conferencesand journals, and more fundamental research publication venues.

We limited our search to research papers published between 2010and 2019 to focus on recent advances in the field. We included allpublication venues as we wished to be inclusive of all CV researchersfocusing on problems in the Global South. We consider inclusionto be critical in achieving a thorough understanding of the manydifferent needs addressed by CV researchers, and the ethical risks ofCV research in the Global South.

To identify appropriate search queries, we first began with anexploratory search on ACM Digital Library. We defined our searchqueries as two components: one CV term and one contextual term,e.g. “smart camera” AND “Global South”. To build our initiallist of search queries, we selected as CV terms common wordsand phrases used in CV literature, e.g. “camera”, “drone” and“machine vision”. As contextual terms, we chose common wordsreferring to HCI4D/ICTD research areas and locations, and ethical

0

2

4

6

8

10

12

14

16

18

2010 2011 2012 2013 2014 2015 2016 2017 2018 2019

Figure 1: Distribution of publication years for the research papers inour corpus (n = 55).

risks in technology , e.g. “privacy”, “development”, “poverty” and“Bangladesh”. Our initial list consisted of 15 CV terms and 24contextual terms, resulting in 360 search queries.

We selected papers based on their title, looking only at the first100 results (ranked using ACM DL’s relevance criterion). We thendiscarded terms based on the relevancy of the selected papers.Terms that were too general and returned too many irrelevant articles(e.g. CV term “mobile phone” or the contextual term “develop-ment”) or that continuously returned no relevant content (e.g. CVterm “webcam” or contextual term “racism”) were discarded.

We retained 5 CV terms (Computer Vision, Drone, SmartCamera, Machine Vision, Face Recognition) and 10 contextualterms (ICTD, ICT4D, Global South, Surveillance, Ethics, Pakistan,Bangladesh, India, Africa, Ethiopia) for a total of 50 search queriesper platform. This selection allowed us to efficiently cover a widespectrum of CV research areas (CV terms) focused on Global Southcountries (contextual terms). Additionally, this also provided uswith an alternative way of covering relevant work by focusing onthose considering risks and ethics (contextual terms) in CV.

A total of 470 papers were collected over all four platforms usingour defined search queries. We then read attentively each abstractand skimmed through each paper to evaluate their relevancy usingour aforementioned criterion. Ultimately, 55 unique research paperswere deemed relevant.

For each paper, we noted the publication year, the country theCV system was designed for (target location), and the country (re-searchers’ location) and region (Global South or Global North) ofthe researchers who designed the systems (i.e. the country wherethe authors’ institution is located2). Additionally, we noted the maintopic of the visual data used by the system (e.g. cars), the applicationarea (i.e. the addressed needs, e.g. healthcare), and the ethical risksat play defined in Sect. 2. For all risks, we only considered likelyoccurrences, i.e. cases where a system appears to exhibit certaindirect risks following normal use.

We did not collect more fine-grained characteristics such as pre-cise deployment areas in our data collection process due to theeclecticism of our corpus: deployments, if they occurred, were notalways discussed in detail.

The first author then performed open and axial coding [17] onboth the application areas and data topics to extract the principalcategories best representing the corpus.

4 RESULTS

In this section, we describe the characteristics of the research paperscollected (n = 55) for our analysis. Table 1 shows an excerpt of ourfinal data collection table used to generate the results below. Thecomplete table is included in the supplemental material. Publicationyears are shown in Fig. 1.

2For papers with authors from different countries, we noted the locationmost shared by the authors.

4.1 Location

Researchers’ location. Of the 55 articles, 41 were authored byresearchers from the Global South and 14 by researchers from theGlobal North. Specifically, Global South researchers were locatedin India (20), Bangladesh (10), Pakistan (6), Brazil (1), Iran (1),Iraq (1), Kenya (1), and Malaysia (1). Global North researchers werelocated in US (11), Germany (1), Japan (1), and South Korea (1).

Target location. For papers authored by Global South re-searchers, the systems were designed for, or deployed in the samecountry as the researchers’ institutions. For the 14 systems de-signed by Global North researchers, the target locations were varied:Ethiopia (2), India (2), Kenya (1), Mozambique (1), Pakistan (1),Senegal (1), and Zimbabwe (1). We also note five systems designedto cover broader areas: Global South (4), and Africa (1).

4.2 Data topics

Using open and axial coding [17], we extracted six broad categoriescovering all visual data topics found in our corpus: characters, hu-man, medicine, outdoor scenes, products, and satellite. Each paperis labeled with only one category: the category that fits best with thespecific data topic of the paper.

Characters. This category encompasses any visual data thatcan be used as the input of an optical character recognition system.Systems using text [64], license plates [38, 43], checks [7], andbanknotes [55] fit into this category.

Human. Included in this category are any visual data wherehumans are the main subjects. Such data is typically used inface recognition [12, 39], sign recognition [36, 65], tracking [67],detection [66, 70], and attribute estimation [33]. In our corpus,faces [12, 31, 39] and actions [36, 65, 66, 70] represent the mainobjects of study using such data.

Medicine. Medical data is tied to the study of people’s health.We include in this category any image or video of people with afocus on healthcare-related attributes [47], medical documents [21],and diagnostic test images [20, 22].

Outdoor scenes. Any visual data in which the main focus ison outdoor scenes is grouped under this category. This includes datacollected by humans [51], but also cameras mounted on vehicles[11, 18, 79], and surveillance cameras [23, 49]. We however excludeairborne and satellite imagery (i.e. high-altitude, top-down views)from this category.

Products. We group in this category visual data focusing onobjects resulting from human labour and fabrication processes suchas fruit [30, 68], fish [37, 72], silk [63], and leather [24].

Satellite. Airborne and satellite imagery data such as multi-spectral imagery [59], are grouped under this category.

The distribution of these topics in our corpus is shown in Fig. 2.Fig. 3 shows how this distribution varies with the researchers’ region.

4.3 Application areas

Through the use of open and axial coding [17], we identified 7 broadapplication area categories representing all applications found inour corpus: agriculture and fishing, assistive technology, healthcare,policy planning, safety, surveillance, and transportation.

Agriculture and fishing. We include in this category applica-tions related to agriculture, fishing, and industrial uses of goods fromthese areas. For example, systems were designed to automaticallyrecognize papaya diseases [30], identify fish exposed to heavy met-als [37], and estimate crop areas in Ethiopia using satellite data [56].

Title Year TargetLocation

Researchers’Location(Region)

Data Topic ApplicationAreas Risks

Cross border intruder detection in hilly ter-rain in dark environment [66] 2016 India India

(Global South) Human Policy plan.,Surveillance

Discrimination,Privacy,Spoofing

Field evaluation of a camera-based mobilehealth system in low-resource settings [22] 2014 Zimbabwe United States

(Global North) Medicine Healthcare N/A

Table 1: Excerpt from the final data collection log used in this study. The complete table is included in the supplemental material.

0 5 10 15

Satellite

Products

Outdoor scenes

Medicine

Human

Characters

Figure 2: Distribution of the visual data topics (n = 55) in our corpus(n = 55).

0 0.1 0.2 0.3 0.4 0.5

Satellite

Products

Outdoor scenes

Medicine

Human

Characters

Figure 3: Normalized distributions of data topics for systems designedby researchers from the Global North (blue, n = 14) and from theGlobal South (orange, n = 41) .

Assistive technology. Research papers in this category studythe design of CV systems assisting people with disabilities in doingeveryday tasks. This includes Bengali sign language recognitionsystems [36, 65], the Indian Spontaneous Expression Database usedfor emotion recognition [31], and a Bangla currency recognitionsystem to assist people with visual impairments [55].

Healthcare. CV research focused on healthcare-related topicsfall into this category. We note adaptive games for stroke rehabilita-tion patients in Pakistan using CV technology to track players [40].We also include systems used to analyze diagnostic tests [22], anddigitize medical forms [21].

Policy planning. We include in this category policies tied tourban planning, development planning, environment monitoring,traffic control, and general quality-of-life improvements. We noteSpotGarbage used to detect garbage in the streets of India [51],a visual pollution detector in Bangladesh [3], and various trafficmonitoring systems in Kenya [41] and Bangladesh [23, 60]. We alsohighlight uses of satellite data to estimate the poorest villages inKenya to guide cash transfers towards people in need [1], and togain a better understanding of rural populations of India [34].

Safety. In the context of our work, we define safety as theprotection of people from general harms. We include in this cate-gory CV systems for bank check fraud detection [7], food toxicitydetection [37], and informal settlement detection [28].

0 5 10 15 20

Transportation

Surveillance

Safety

Policy Plan.

Healthcare

Assistive Tech.

Agri. & Fishing

Figure 4: Distribution of application area categories (n = 80) in ourcorpus (n = 55).

0 0.1 0.2 0.3 0.4 0.5

Transportation

Surveillance

Safety

Policy Plan.

Healthcare

Assistive Tech.

Agri. & Fishing

Figure 5: Normalized distributions of application areas for systemsdesigned by researchers from the Global North (blue, n = 23) andfrom the Global South (orange, n = 57).

Surveillance. Surveillance applications refer to systems moni-toring people or activities to protect a specific group from danger.We make no distinction between explicit and secret acts of surveil-lance, i.e. being monitored without one’s consent. We found systemsfor theft detection [70], border intrusion detection in India [66], au-tomatic license plate recognition [38], and tracking customers [67].

Transportation. CV systems in this category focus on improv-ing and understanding transport conditions in the Global South. Themajority of the systems identified in our corpus were aimed at moni-toring and reducing traffic [49, 60] and accidents [23]. We also notesystems addressing the problem of autonomous driving [11, 79].

Discrim

inatio

n

Privacy

Psych

. harm

s

Security

breach

es

Spo

ofin

g

Agri. & Fishing 0 1 0 0 6

Assistive Tech. 1 1 1 0 2

Healthcare 0 0 0 0 0

Policy Plan. 4 14 0 0 12

Safety 1 7 0 0 8

Surveillance 9 15 10 0 14

Transportation 0 9 3 0 9

Agri. &

Fishin

g

Assistive Tech

.

Health

care

Po

licy Plan

.

Safety

Surveillan

ce

Transp

ortatio

n

Characters 0 2 0 0 1 3 3

Human 0 3 2 2 0 9 0

Medicine 0 0 4 0 0 0 0

Outdoor scenes 0 0 0 7 4 2 8

Products 12 0 0 1 1 0 0

Satellite 1 0 0 9 5 1 0

Figure 6: Co-occurrence frequencies between data topics and appli-cation areas in our corpus (n = 55).

0 5 10 15 20 25 30 35

Spoofing

Security breaches

Psych. harms

Privacy

Discrimination

Figure 7: Distribution of risks (n = 80) in our corpus (n = 55). 18papers had no identified risks.

0 0.1 0.2 0.3 0.4 0.5

Spoofing

Security breaches

Psych. harms

Privacy

Discrimination

Figure 8: Normalized distributions of risks (n = 80) for systems de-signed by researchers from the Global North (blue, n = 16) and fromthe Global South (orange, n = 64).

Each paper was tagged with at least one category and at mostthree categories (median = 1). The distribution of application areasis shown in Fig. 4, followed by the normalized distributions for eachresearchers’ region in Fig. 5. We present the global co-occurrencefrequencies between data topics and application areas in Fig. 6.

4.4 Risks

We briefly illustrate how we assessed the presence of risks in CVsystems (Sect. 3), using two examples from our corpus. Note that wedid not consider intent in our analysis; instead, we focused strictlyon the possibility of enabling certain risks.

First, we consider a system using satellite imagery designed toidentify informal settlements in the Global South, informing NGOsof regions that could benefit the most from their aid. [28]. While thesystem focuses on areas rather than individuals, it nevertheless col-lects information on vulnerable communities without their consent(privacy violations). Additionally, a malicious actor aware of thesystem’s functionalities could modify settlements to hide them insatellite imagery, e.g. by covering rooftops with different materials(spoofing). For these reasons, we tagged the system with two risks.

Second, we examine a sign language digit recognition system [36],which we tagged with no risk. As the system appears to solely relyon hand gestures and focuses on a simple limited task, the presenceof all 5 risks appeared minimal.

Fig. 7 shows the overall risk distribution for our corpus. Paperswere tagged with at most four risks; 18 had no risks (median =1). Normalized distributions for each researchers’ region appearin Fig. 8. We show the global co-occurrence frequencies betweenapplication areas and risks in Fig. 9, and between data topics andrisks in Fig. 10. Finally, we present the conditional probability of arisk being present, given the presence of another in Fig. 11.

Discrim

inatio

n

Privacy

Psych

. harm

s

Security

breach

es

Spo

ofin

g

Agri. & Fishing 0 1 0 0 6

Assistive Tech. 1 1 1 0 2

Healthcare 0 0 0 0 0

Policy Plan. 4 14 0 0 12

Safety 1 7 0 0 8

Surveillance 9 15 10 0 14

Transportation 0 9 3 0 9

Agri. &

Fishin

g

Assistive Tech

.

Health

care

Po

licy Plan

.

Safety

Surveillan

ce

Transp

ortatio

n

Characters 0 2 0 0 1 3 3

Human 0 3 2 2 0 9 0

Medicine 0 0 4 0 0 0 0

Outdoor scenes 0 0 0 7 4 2 8

Products 12 0 0 1 1 0 0

Satellite 1 0 0 9 5 1 0

Figure 9: Co-occurrence frequencies between application areas andrisks in our corpus (n = 55).

Discrim

inatio

n

Privacy

Psych

. harm

s

Security

bre

ache

s

Spo

ofin

g

Characters 0 3 3 0 5

Human 10 10 7 0 9

Medicine 0 0 0 0 0

Outdoor scenes 0 6 0 0 8

Products 0 0 0 0 5

Satellite 2 7 0 0 5

Figure 10: Co-occurrence frequencies between data topics and risksin our corpus (n = 55).

Discrim

inatio

n

Privacy

Psych

. harm

s

Security

breach

es

Spo

ofin

g

Characters 0 3 3 0 5

Human 10 10 7 0 9

Medicine 0 0 0 0 0

Outdoor scenes 0 6 0 0 8

Products 0 0 0 0 5

Satellite 2 7 0 0 5

Discrim

inatio

n

Privacy

Psych

. harm

s

Security

breach

es

Spo

ofin

g

Discrimination 1 0.46 0.70 - 0.34

Privacy 1 1 1 - 0.66

Psych. Harms 0.58 0.38 1 - 0.28

Security breaches 0 0 0 - 0

Spoofing 0.92 0.81 0.90 - 1

Figure 11: Conditional probability of a risk being present, given thepresence of another risk p(row|column) in our corpus (n = 55).

5 DISCUSSION

Overall, we find that the majority of surveyed CV systems were de-signed to address specific needs of a community, ranging from foodquality control to healthcare, traffic minimization, and improvedsecurity. Moreover, we observe that currently popular researchareas in the West, such as autonomous vehicles and affective com-puting, have not achieved the same penetration level so far in theGlobal South.

We present below our most notable findings and compare themwith the findings of Skirpan and Yeh for the Global North [74]. Wethen present design considerations to minimize each risk, followedby a discussion on the limitations of our study.

5.1 Key findingsCV research for the Global South, by the Global South?

75% of the papers in our corpus were authored by researchers fromthe Global South. We find this proportion concerning, as it highlightsan ever present risk of colonization of methods and knowledge [75].This is furthermore important as both the interpretation of visual dataand its associated applications are highly dependent on the contextin which the data is situated [25]. Thus, we believe this proportionraises an important ethical concern around the interpretation of visualdata and the politics tied to the downstream CV applications [35].

Interests differ between regional research groups.Grouping the distributions of data topics and application areas byresearchers’ regions (Fig. 3 and 5) reveals significant differences ininterests. We note for example the data topic of satellite imageryused in 50% of all Global North systems and only 5% of GlobalSouth systems. We speculate that this type of data may be lessutilized by Global South researchers as it is typically used to solvelarger fundamental problems rather than directly addressing specificneeds. Conversely, we find Global North researchers to focus moreon data that can be used to solve broader problems.

These distinctions are also visible in application areas: GlobalNorth researchers appear to focus much more on applications re-quiring significant technical expertise, often employing modalitiesneither aligned with nor originated from local methods [75], withbroader ramifications (e.g. policy planning applications identify-ing at-risk areas through satellite imagery [1, 28]) than applicationsfocused on specific local problems, requiring a thorough understand-ing of the community and less technical expertise (e.g. applicationsfor transportation, agriculture, and surveillance). In contrast, theregional distribution for Global South researchers is much moreuniform, with greater interest in those local problems. Furthermore,these disparities may also be caused by the lack of deeper knowl-edge about the local problems of the Global South among the CVcommunity in the Global North [35].

Finally, we note that such differences in interests may also becaused by accessibility issues, e.g. data being made available onlyto certain researchers, high cost to developing new system comparedto imports, and computational power issues.

Surveillance applications – promising, but not without risks.Globally, we find surveillance to be the second most popular appli-cation area behind policy planning, an area mostly popular due to itsencompassing definition. We attribute this popularity most notablyto one factor: surveillance systems are a natural solution to satisfyin part the essential need of security, albeit at the cost of restrictingliberties. This trade-off is notably closely tied to the government sys-tems in place, which vary wildly in the Global South [6]. Moreover,we note that the motivations for surveillance vary between cultures,ranging from care, nurture, and protection to control and oppression.

We emphasize that researchers should be especially careful indesigning surveillance systems, as we found that most studied herecould be easily misused to focus on increasing control over peoplerather than improving security; we found surveillance to be theapplication most frequently tied to any risk, and the one with themost diverse set of risks (Fig. 9).

The impact of machine learning on CV risks.73% of all CV systems analyzed used machine learning techniquesto achieve their goals. This has several implications when it comesto the presence of certain risks in CV systems. First, it amplifiesthe risk of spoofing and adversarial inputs, the most frequent risk inour corpus appearing in nearly all data topics and application areas(Fig. 9 and 10). Indeed, research has shown neural networks to beeasily misguided by adversarial inputs [27, 57, 61].

Second, the use of machine learning amplifies the risk of dis-crimination, in particular the unplanned type, i.e. CV systems usingsensitive attributes for decision making without being designed todo so. We found fewer examples of this risk in our corpus as dis-crimination occurs strictly against people; the risk was mainly foundin surveillance and policy planning applications using human data.

Privacy violations – the common denominator.In our analysis, we found privacy violations to be the second mostcommon risk, far ahead discrimination and psychological harms.Specifically, we found the risk of privacy violations most present insurveillance and policy planning applications, both areas requiringin-depth knowledge of people.

Additionally, we have observed that certain risks were typicallypresent only when privacy violations were too. In particular, therisk of privacy violations was always present when the risks ofdiscrimination and psychological harms were found (Fig. 11). Whileour analysis does not robustly assess the relations between eachrisk, this finding suggests an almost hierarchical relation betweencertain risks, beginning with privacy violations: e.g, for automateddiscrimination to occur, sensitive attributes of people must first beknown (privacy violations). Similarly, the risk of psychologicalharms depends directly on the presence of other risks. We notehowever that the risk of spoofing appears to be tied more closely tocertain algorithms than risks, although we still expect its frequencyto rise as access to sensitive data (privacy violations) increases.

Privacy violations frequency – a likely underestimate.We note that research on privacy and technology [2,6] has shown pri-vacy to be regionally constructed, differing between cultural regions:e.g. people in the Global South often perceive privacy differentlythan Western liberal people. Our interpretation of the general defini-tion of privacy violations used in our analysis (Sect. 2) likely doesnot encompass every cultural definition of privacy in the GlobalSouth. As such, there may exist many unforeseen privacy-relatedrisks tied to CV systems, leading us to believe the actual frequencyof the risk is even higher than what we have found in our analysis.

Not all risks could be reliably perceived.We note in our analysis two risks that could not be reliably estimated:security breaches and psychological harms. In fact, we were unableto find a single case of the former. We attribute this primarily tothe method used for analysis: detection of the risk would requireresearch papers to present an in-depth description of a system’s im-plementation, highlighting certain vulnerabilities. Similarly, whilewe have found some systems with risks of psychological harms, weemphasize that we have likely underestimated its actual frequency:detection requires both an in-depth understanding of the system andits subsequent deployments. Moreover, compared to the other fourrisks, we re-emphasize that the risk of psychological harms resultsfrom longer-term interactions and may only become discernible asrisk-driven harms begin to occur.

Frequency versus severity.We first stress that the absence of certain risks does not indicate theirnon-existence. On the contrary, this informs us to be more waryof these risks as we may have little means to diagnose them in thiscontext. Additionally, we note that the risk frequency estimated hereis a different concept than the frequency of harm occurrences tied tothe aforementioned risk. For example, while spoofing was found tobe the most common risk in our corpus, harm occurrences linked tothis risk requires the existence of a knowledgeable malicious actorattacking the system, which may be much more uncommon.

Furthermore, we emphasize that each risk can have severe im-pacts, regardless of their frequency. For example, while the risk ofsecurity breaches may be infrequent, a single occurrence could leadto disastrous consequences, from large-scale information leaks tocasualties. Thus, while this study may highlight common pitfalls tiedto the design of recent CV systems, we urge designers to considerthe potential impacts caused by all risks equally.

5.2 Comparison with Global North risk analysis5.2.1 Risk probabilityIn their work, Skirpan and Yeh [74] identified discrimination, secu-rity breaches and psychological harms as the most probable risks. Inthe present study, we instead identify privacy violations and spoofingas the most probable risks in the Global South. We attribute thesedifferences to two reasons.

First, each work focused on a different body of work. Skirpan andYeh searched through both research literature and current news, lead-ing to a more diverse set of systems and problems. We, on the other

hand, limited our search to research papers focusing specifically onthe Global South to obtain a more grounded understanding. Thus,risks that cannot be reliably estimated from research papers (e.g.security breaches) did not appear frequently in our study. However,this alone is insufficient to suggest the risk space to be different inthe Global South: the absence of a risk in research literature doesnot imply its non-existence in the real world.

Second, we believe these differences to be primarily related to theavailability and the acceptance of CV systems in both regions. CVsystems in the West, which have become nearly ubiquitous, are oftendesigned to be considerate of the history and politics of the region.However, in the Global South, fewer systems are designed and, inmost cases, they involve transferring existing Northern technologiesto the Global South. The designed systems thus tend to be less con-siderate of the culture and politics of the Global South [35], limitingamong other things their level of adoption by the community. More-over, we re-emphasize the pseudo-hierarchical nature of the risksstudied here: privacy violations appear to serve as a springboardfor the other risks to arise. Thus, in areas where CV technology isrelatively recent, we should expect to find privacy violations to bemore frequent than other risks. This, combined with the extensiveuse of machine learning techniques, can also explain why the risk ofspoofing and adversarial inputs was found to be more frequent thanthe three risks higlighted by Skirpan and Yeh.

5.2.2 Risk PerceptionSkirpan and Yeh [74] have also evaluated risks in terms of severityand uncertainty, leading them to identify discrimination and securitybreaches as the most potential important risks. While we have notformally performed this type of analysis in our work, we speculatethe problem space in the Global South to be different.

We believe that at this time, the most important risk is privacyviolations, in part due to its gateway property to other risks and thecurrent availability of CV systems in the Global South. We noteprivacy violations can not only enable automated discrimination inCV systems, but also broader discrimination and security risks. Forexample, by surveilling and collecting specific sensitive attributesof its people, a state could identify and arrest, displace, or mistreatcertain marginal groups [53].

Additionally, our perception aligns with a growing body ofHCI4D/ICTD research on the more general problem of privacyand technology [2, 4–6, 32, 73] showing privacy to be regionallyconstructed and highlighting a general lack of understanding fromtechnology designers, which in turn suggests there may be importantprivacy (and security) concerns notably at gender and communitylevels unforeseen in the Global North.

5.3 Design considerations for risk minimizationAs a general consideration to minimize risks in CV systems, weurge researchers to ensure their designs are context-appropriate forthe focused area. Designing such systems requires expertise intechnical domains such as CV and design, but also in humanities andsocial sciences to identify potential risks for a given community. Assuch, we recommend involving more local researchers in the designprocess. Aside from risk minimization, the participation of localpeople is critical in ensuring the goals of the system are aligned withthe true needs of the community. We refer designers to Irani et al.’swork on postcolonial computing [35] for more in-depth guidance onthis design process. For the rest of this section, we provide designconsiderations for each ethical risk considered in our study.

Privacy violations. We briefly highlight two common privacy-related issues in modern CV systems. First, decision making pro-cesses have become increasingly obfuscated: decisions could un-knowingly be based on private attributes, rather than the ones ex-pected by the designer and consented by the user. Second, non-usersin shared spaces may unconsciously be treated as users. We refer

to a growing body of work on privacy-preserving CV for concretesolutions [16, 69, 71, 76], such as Ren et al.’s face anonymizer foraction detection. [69].

Discrimination. Discrimination in smart CV systems and gen-eral AI systems has been extensively researched in recent years[15, 83]. The use of biased, imbalanced, or mis-adapted data andtraining schemes are typically the cause of this risk. We present twostrategies to address these issues. First, datasets can be adapted andbalanced by collecting new context-appropriate data, e.g. ensuringthat every gender and ethnicity is equally represented in a humandataset [19]. Second, learning schemes can be modified to achievecertain definitions of fairness [10, 52, 82]. We refer to Barocas etal.’s book [13] for a more in-depth analysis of general fairness-basedsolutions to minimize discrimination.

Spoofing and adversarial inputs. In general, risk minimiza-tion can be achieved by improving system robustness. Solutionsinclude taking into account external conditions that could affect per-formance (e.g. variable lighting), using sets of robust discriminativefeatures in decision making, and improving measurement’s qual-ity. However, all systems are not equally prone to fall to the sametypes of attacks. For example, systems involving neural networksare typically weaker to precise adversarial inputs [27, 57, 61]. Inthis context, we refer to recent advances in the field for specificsolutions [8, 9, 29, 46].

Security breaches. Our considerations here are limited as theminimization of this risk depends heavily on the type of system. Westrongly encourage designers to engage with experts in this fieldto develop robust context-specific solutions. Nevertheless, generalsolutions include avoiding system deployments in areas with easilyaccessible sensitive information and limiting access to the system.

Psychological harms. We urge designers to engage with thecommunities and involve local experts to determine what constitutesan acceptable and helpful CV system, and to follow these guidelinesduring design and deployment. We again emphasize how this risk isclosely tied to all four others: any occurrence of harm related to theother risks will inevitably lead people to modify their lifestyle. Thus,the minimization of this risk depends directly on the minimizationof the other four risks.

5.4 LimitationsData collection. Our analysis does not take into account every

use of CV in the Global South. Indeed, we have only consideredcertain systems that were published as research papers in specificdigital libraries and tagged as relevant during our search. We thusonly present a partial view of the real practical uses of CV in theGlobal South, i.e. systems encountered by the public in their dailylives, and the subsequent ethical risks.

Furthermore, we note that our relevance criterion can be inter-preted and applied differently. For example, “use of visual datarelated to the area” could refer to the use of a publicly-availabledataset representative of the target location. However, it could alsorefer to any imagery that appears to be related to the area. Due to thescarcity of such datasets, we opted for the latter definition, whichresulted in a more relaxed definition of relevancy.

Finally, we acknowledge that the collected characteristics arefor the most part surface-level, in part due to the eclecticism of thecorpus: e.g. precise deployment areas (e.g. office spaces) could notbe collected uniformly as deployments were not always discussed.

Risk evaluation. First, we were unable to reliably assess thepresence of the risks of security breaches and psychological harmsin our corpus. While we have attempted to mitigate this issue byestimating the risks’ frequency based on our understanding of thesystems, our evaluation does not likely represent well their actualpresence in the Global South.

Second, we note that our corpus was heavily skewed towardscertain Global South countries: 69% of the systems were designedfor the countries of India, Bangladesh and Pakistan. We thus ac-knowledge that our results do not represent equally well, and maynot generalize well to every community in the Global South.

Finally, we note that our risk frequency and perception analysesdirectly reflect our views of the problem, namely as designers andresearchers from the Global North, which are themselves inspiredby previous work [74]. Nevertheless, we believe that our overviewof the typical CV-related ethical risks can serve as a basis to assistresearchers in designing safe context-appropriate CV systems forthe Global South.

6 CONCLUSION

We have presented an overview of applications and ethical risks tiedto recent advances in CV research for the Global South. Our resultsindicate that surveillance and policy planning are the most exploredapplication areas. We found research interests to differ notablybetween regions, with researchers from the North focusing more onbroader problems requiring complex technical solutions. Risk-wise,privacy violations appears as the most likely and most severe risk toarise in CV systems. This last result differs from those of Skirpan andYeh [74], which were obtained by analyzing both research literatureand news events with a focus on the Global North. Taken together,our findings suggest that the actual uses of CV and the importanceof its associated ethical risks are region-specific, depending heavilyon a community’s needs, norms, culture and resources.

As future work, this study can be extended to non-research CVapplications, e.g. commercial systems, to gain a more thoroughunderstanding of the situation in the Global South. Additionally,similarly to what Skirpan and Yeh proposed in their work [74], webelieve the next major step in improving our understanding of theseissues is to survey direct and indirect users of CV systems and studytheir interactions in specific areas of the Global South. Only thencan we gain a more precise and representative understanding of theimportance and the frequency of each of those ethical risks.

REFERENCES

[1] B. Abelson, K. R. Varshney, and J. Sun. Targeting direct cash transfersto the extremely poor. In Proc. of the 20th ACM SIGKDD InternationalConference on Knowledge Discovery and Data Mining, KDD ’14, pp.1563–1572. ACM, NY, USA, 2014. doi: 10.1145/2623330.2623335

[2] N. Abokhodair and S. Vieweg. Privacy & social media in the contextof the arab gulf. In Proc. of the 2016 ACM Conference on DesigningInteractive Systems, DIS ’16, p. 672–683. ACM, NY, USA, 2016. doi:10.1145/2901790.2901873

[3] N. Ahmed, M. N. Islam, A. S. Tuba, M. Mahdy, and M. Sujaud-din. Solving visual pollution with deep learning: A new nexus inenvironmental management. Journal of Environmental Management,248:109253, October 2019. doi: 10.1016/j.jenvman.2019.07.024

[4] S. Ahmed, M. Haque, J. Chen, and N. Dell. Digital privacy chal-lenges with shared mobile phone use in bangladesh. Proc. of theACM on Human-Computer Interaction, 1(CSCW), 11 2017. doi: 10.1145/3134652

[5] S. I. Ahmed, S. Guha, M. R. Rifat, F. H. Shezan, and N. Dell. Privacy inrepair: An analysis of the privacy challenges surrounding broken digitalartifacts in bangladesh. In Proc. of the Eighth International Conferenceon Information and Communication Technologies and Development,ICTD ’16. ACM, NY, USA, 2016. doi: 10.1145/2909609.2909661

[6] S. I. Ahmed, M. R. Haque, S. Guha, M. R. Rifat, and N. Dell. Privacy,security, and surveillance in the global south: A study of biometricmobile sim registration in bangladesh. In Proc. of the 2017 CHIConference on Human Factors in Computing Systems, CHI ’17, p.906–918. ACM, NY, USA, 2017. doi: 10.1145/3025453.3025961

[7] Y. Akbari, M. J. Jalili, J. Sadri, K. Nouri, I. Siddiqi, and C. Djeddi. Anovel database for automatic processing of persian handwritten bankchecks. Pattern Recognition, 74(C):253 – 265, February 2018. doi: 10.1016/j.patcog.2017.09.011

[8] N. Akhtar, J. Liu, and A. Mian. Defense against universal adversarialperturbations. In The IEEE Conference on Computer Vision and PatternRecognition (CVPR), pp. 3389–3398. IEEE, June 2018. doi: 10.1109/CVPR.2018.00357

[9] N. Akhtar and A. Mian. Threat of adversarial attacks on deep learningin computer vision: A survey. IEEE Access, 6:14410–14430, 2018.doi: 10.1109/ACCESS.2018.2807385

[10] L. Anne Hendricks, K. Burns, K. Saenko, T. Darrell, and A. Rohrbach.Women also snowboard: Overcoming bias in captioning models. InThe European Conference on Computer Vision (ECCV), pp. 793–811.Springer, September 2018. doi: 10.1007/978-3-030-01219-9 47

[11] A. Badshah, N. Islam, D. Shahzad, B. Jan, H. Farman, M. Khan,G. Jeon, and A. Ahmad. Vehicle navigation in gps denied environmentfor smart cities using vision sensors. Computers, Environment andUrban Systems, 77:101281, September 2019. doi: 10.1016/j.compen-vurbsys.2018.09.001

[12] S. Banerjee, A. Bhattacharjee, and S. Das. Deep domain adaptationfor face recognition using images captured from surveillance cameras.In 2018 International Conference of the Biometrics Special InterestGroup (BIOSIG), pp. 1–12. IEEE, Sep. 2018. doi: 10.23919/BIOSIG.2018.8553278

[13] S. Barocas, M. Hardt, and A. Narayanan. Fairness and MachineLearning. fairmlbook.org, 2019. http://www.fairmlbook.org.

[14] A. L. Blank. Computer vision machine learning and future-orientedethics. Honors Projects, 107:1–15, 2019.

[15] J. Buolamwini and T. Gebru. Gender shades: Intersectional accuracydisparities in commercial gender classification. In Conference onFairness, Accountability and Transparency (FAT), vol. 81 of Proc. ofMachine Learning Research, pp. 77–91. PMLR, 2018.

[16] J. Chen, J. Konrad, and P. Ishwar. Vgan-based image representationlearning for privacy-preserving facial expression recognition. In TheIEEE Conference on Computer Vision and Pattern Recognition (CVPR)Workshops, June 2018.

[17] J. M. Corbin and A. Strauss. Grounded theory research: Procedures,canons, and evaluative criteria. Qualitative sociology, 13(1):3–21,1990. doi: 10.1007/BF00988593

[18] A. Danti, J. Y. Kulkarni, and P. Hiremath. An image processing ap-proach to detect lanes, pot holes and recognize road signs in indianroads. International Journal of Modeling and Optimization, 2(6):658–662, December 2012. doi: 10.7763/IJMO.2012.V2.204

[19] T. de Vries, I. Misra, C. Wang, and L. van der Maaten. Does objectrecognition work for everyone? In The IEEE Conference on ComputerVision and Pattern Recognition (CVPR) Workshops, June 2019.

[20] N. Dell and G. Borriello. Mobile tools for point-of-care diagnosticsin the developing world. In Proc. of the 3rd ACM Symposium onComputing for Development, ACM DEV ’13, pp. 9:1–9:10. ACM, NY,USA, 2013. doi: 10.1145/2442882.2442894

[21] N. Dell, J. Crawford, N. Breit, T. Chaluco, A. Coelho, J. McCord,and G. Borriello. Integrating ODK scan into the community healthworker supply chain in mozambique. In Proc. of the Sixth Interna-tional Conference on Information and Communication Technologiesand Development, ICTD ’13, pp. 228–237. ACM, NY, USA, 2013. doi:10.1145/2516604.2516611

[22] N. Dell, I. Francis, H. Sheppard, R. Simbi, and G. Borriello. Fieldevaluation of a camera-based mobile health system in low-resourcesettings. In Proc. of the 16th International Conference on Human-computer Interaction with Mobile Devices & Services, MobileHCI’14, pp. 33–42. ACM, NY, USA, 2014. doi: 10.1145/2628363.2628366

[23] M. M. L. Elahi, R. Yasir, M. A. Syrus, M. S. Q. Z. Nine, I. Hossain, andN. Ahmed. Computer vision based road traffic accident and anomalydetection in the context of bangladesh. In 2014 International Confer-ence on Informatics, Electronics Vision (ICIEV), pp. 1–6. IEEE, May2014. doi: 10.1109/ICIEV.2014.6850780

[24] U. Farooq, M. U. Asad, F. Rafiq, G. Abbas, and A. Hanif. Applicationof machine vision for performance enhancement of footing machineused in leather industry of pakistan. In Eighth International Confer-ence on Digital Information Management (ICDIM 2013), pp. 149–154.IEEE, Sep. 2013. doi: 10.1109/ICDIM.2013.6694000

[25] V. Flusser and N. A. Roth. Into the Universe of Technical Images,vol. 32. University of Minnesota Press, ned - new edition ed., 2011.

[26] C. Garvie, A. Bedoya, and J. Frankle. The perpetual line-up: Unregu-lated police face recognition in america. October 2016.

[27] I. Goodfellow, J. Shlens, and C. Szegedy. Explaining and harness-ing adversarial examples. In International Conference on LearningRepresentations, 2015.

[28] B. J. Gram-Hansen, P. Helber, I. Varatharajan, F. Azam, A. Coca-Castro, V. Kopackova, and P. Bilinski. Mapping informal settlementsin developing countries using machine learning and low resolutionmulti-spectral data. In Proc. of the 2019 AAAI/ACM Conference on AI,Ethics, and Society, AIES ’19, pp. 361–368. ACM, NY, USA, 2019.doi: 10.1145/3306618.3314253

[29] C. Guo, M. Rana, M. Cisse, and L. van der Maaten. Countering adver-sarial images using input transformations. In International Conferenceon Learning Representations, pp. 1–12, 2018.

[30] M. T. Habib, A. Majumder, A. Jakaria, M. Akter, M. S. Uddin, andF. Ahmed. Machine vision based papaya disease recognition. Journalof King Saud University - Computer and Information Sciences, 2018.doi: 10.1016/j.jksuci.2018.06.006

[31] S. L. Happy, P. Patnaik, A. Routray, and R. Guha. The indian sponta-neous expression database for emotion recognition. IEEE Transactionson Affective Computing, 8(1):131–142, jan – mar 2017. doi: 10.1109/TAFFC.2015.2498174

[32] S. M. T. Haque, P. Saha, M. S. Rahman, and S. I. Ahmed. Of ulti,“hajano”, and ‘matachetar otanetak datam’: Exploring local practices ofexchanging confidential and sensitive information in urban bangladesh.Proc. ACM Hum.-Comput. Interact., 3(CSCW), November 2019. doi:10.1145/3359275

[33] S. Helleringer, C. You, L. Fleury, L. Douillot, I. Diouf, C. T. Ndiaye,V. Delaunay, and R. Vidal. Improving age measurement in low- andmiddle-income countries through computer vision: A test in senegal.Demographic Research, 40:219–260, January 2019. doi: 10.4054/DemRes.2019.40.9

[34] W. Hu, J. H. Patel, Z.-A. Robert, P. Novosad, S. Asher, Z. Tang,M. Burke, D. Lobell, and S. Ermon. Mapping missing populationin rural india: A deep learning approach with satellite imagery. InProc. of the 2019 AAAI/ACM Conference on AI, Ethics, and Society,AIES ’19, pp. 353–359. ACM, NY, USA, 2019. doi: 10.1145/3306618.3314263

[35] L. Irani, J. Vertesi, P. Dourish, K. Philip, and R. E. Grinter. Postcolonialcomputing: A lens on design and development. In Proc. of the SIGCHIConference on Human Factors in Computing Systems, CHI ’10, p.1311–1320. ACM, NY, USA, 2010. doi: 10.1145/1753326.1753522

[36] S. Islam, S. S. S. Mousumi, A. S. A. Rabby, S. A. Hossain, and S. Abu-jar. A potent model to recognize bangla sign language digits usingconvolutional neural network. Procedia Computer Science, 143:611 –618, 2018. doi: 10.1016/j.procs.2018.10.438

[37] A. Issac, A. Srivastava, A. Srivastava, and M. K. Dutta. An automatedcomputer vision based preliminary study for the identification of aheavy metal (hg) exposed fish-channa punctatus. Computers in Biologyand Medicine, 111:103326, August 2019. doi: 10.1016/j.compbiomed.2019.103326

[38] V. Jain, Z. Sasindran, A. Rajagopal, S. Biswas, H. S. Bharadwaj, andK. R. Ramakrishnan. Deep automatic license plate recognition system.In Proc. of the Tenth Indian Conference on Computer Vision, Graphicsand Image Processing, ICVGIP ’16, pp. 6:1–6:8. ACM, NY, USA,2016. doi: 10.1145/3009977.3010052

[39] I. Kalra, M. Singh, S. Nagpal, R. Singh, M. Vatsa, and P. B. Sujit.Dronesurf: Benchmark dataset for drone-based face recognition. In2019 14th IEEE International Conference on Automatic Face GestureRecognition (FG 2019), pp. 1–7. IEEE, May 2019. doi: 10.1109/FG.2019.8756593

[40] H. A. Khan, M. Parvez, S. Shahid, and A. Javaid. A comparative studyon the effectiveness of adaptive exergames for stroke rehabilitationin pakistan. In Extended Abstracts of the 2018 CHI Conference onHuman Factors in Computing Systems, CHI EA ’18, pp. LBW044:1–LBW044:6. ACM, NY, USA, 2018. doi: 10.1145/3170427.3188523

[41] A. Kinai, R. E. Bryant, A. Walcott-Bryant, E. Mibuari, K. Welde-mariam, and O. Stewart. Twende-twende: A mobile application fortraffic congestion awareness and routing. In Proc. of the 1st Inter-national Conference on Mobile Software Engineering and Systems,

MOBILESoft 2014, pp. 93–98. ACM, NY, USA, 2014. doi: 10.1145/2593902.2593926

[42] E. Kocabey, M. Camurcu, F. Ofli, Y. Aytar, J. Marin, A. Torralba, andI. Weber. Face-to-bmi: using computer vision to infer body mass indexon social media. In Eleventh International AAAI Conference on Weband Social Media. AAAI Press, 2017.

[43] T. Kumar, S. Gupta, and D. S. Kushwaha. An efficient approach forautomatic number plate recognition for low resolution images. In Proc.of the Fifth International Conference on Network, Communication andComputing, ICNCC ’16, pp. 53–57. ACM, NY, USA, 2016. doi: 10.1145/3033288.3033332

[44] J. Larson, S. Mattu, L. Kirchner, and J. Angwin. How we analyzed thecompas recidivism algorithm — propublica, 2016.

[45] M. Lauronen. Ethical issues in topical computer vision applications.November 2017.

[46] F. Liao, M. Liang, Y. Dong, T. Pang, X. Hu, and J. Zhu. Defense againstadversarial attacks using high-level representation guided denoiser. InThe IEEE Conference on Computer Vision and Pattern Recognition(CVPR), pp. 1778–1787. IEEE, June 2018. doi: 10.1109/CVPR.2018.00191

[47] B. C. Loh and P. H. Then. Mobile Imagery EXchange (MIX) toolkit:Data sharing for the unconnected. Personal and Ubiquitous Computing,19(3–4):723–740, July 2015. doi: 10.1007/s00779-015-0835-2

[48] K. Lum and W. Isaac. To predict and serve? Significance, 13(5):14–19,October 2016. doi: 10.1111/j.1740-9713.2016.00960.x

[49] K. Maeda, T. Morimura, T. Katsuki, and M. Teraguchi. Frugal signalcontrol using low resolution web-camera and traffic flow estimation.In Proc. of the 2014 Winter Simulation Conference, WSC ’14, pp.2082–2091. IEEE, Piscataway, NJ, USA, 2014. doi: 10.5555/2693848.2694108

[50] M. McFarland. Terrorist or pedophile? This start-up says it can outsecrets by analyzing faces. The Washington Post, May 2016.

[51] G. Mittal, K. B. Yagnik, M. Garg, and N. C. Krishnan. Spotgarbage:Smartphone app to detect garbage using deep learning. In Proc. of the2016 ACM International Joint Conference on Pervasive and Ubiqui-tous Computing, pp. 940–945. ACM, NY, USA, 2016. doi: 10.1145/2971648.2971731

[52] A. Morales, J. Fierrez, and R. Vera-Rodrıguez. Sensitivenets: Learningagnostic representations with application to face recognition. arXiv,arXiv/1902.00334, 2019.

[53] P. Mozur. One month, 500,000 face scans: How china is using a.i. toprofile a minority. The New York Times, April 2019.

[54] P. Mozur, J. M. Kessel, and M. Chan. Made in china, exported to theworld: The surveillance state. The New York Times, April 2019.

[55] H. Murad, N. I. Tripto, and M. E. Ali. Developing a bangla currencyrecognizer for visually impaired people. In Proc. of the Tenth Interna-tional Conference on Information and Communication Technologiesand Development, pp. 56:1–56:5. ACM, NY, USA, 2019. doi: 10.1145/3287098.3287152

[56] C. S. Neigh, M. L. Carroll, M. R. Wooten, J. L. McCarty, B. F. Powell,G. J. Husak, M. Enenkel, and C. R. Hain. Smallholder crop areamapped with wall-to-wall worldview sub-meter panchromatic imagetexture: A test case for tigray, ethiopia. Remote Sensing of Environment,212:8 – 20, 2018. doi: 10.1016/j.rse.2018.04.025

[57] A. Nguyen, J. Yosinski, and J. Clune. Deep neural networks are easilyfooled: High confidence predictions for unrecognizable images. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR),pp. 427–436. IEEE, 2015.

[58] S. J. Oh, R. Benenson, M. Fritz, and B. Schiele. Faceless person recog-nition: Privacy implications in social media. In European Conferenceon Computer Vision, pp. 19–35. Springer, 2016. doi: 10.1007/978-3-319-46487-9 2

[59] B. Oshri, A. Hu, P. Adelson, X. Chen, P. Dupas, J. Weinstein, M. Burke,D. Lobell, and S. Ermon. Infrastructure quality assessment in africausing satellite imagery and deep learning. In Proc. of the 24th ACMSIGKDD International Conference on Knowledge Discovery &Data Mining, KDD ’18, pp. 616–625. ACM, NY, USA, 2018. doi: 10.1145/3219819.3219924

[60] T. Osman, S. S. Psyche, J. M. S. Ferdous, and H. U. Zaman. Intelligenttraffic management system for cross section of roads using computer

vision. In 2017 IEEE 7th Annual Computing and CommunicationWorkshop and Conference (CCWC), pp. 1–7. IEEE, Jan 2017. doi: 10.1109/CCWC.2017.7868350

[61] N. Papernot, P. McDaniel, S. Jha, M. Fredrikson, Z. B. Celik, andA. Swami. The limitations of deep learning in adversarial settings. InIEEE European Symposium on Security and Privacy (EuroS P), pp.372–387. IEEE, 2016.

[62] J. Pierce. Smart home security cameras and shifting lines of creepiness:A design-led inquiry. In Proc. of the 2019 CHI Conference on HumanFactors in Computing Systems, CHI ’19. ACM, NY, USA, 2019. doi:10.1145/3290605.3300275

[63] P. Prasobhkumar, C. Francis, and S. S. Gorthi. Automated qualityassessment of cocoons using a smart camera based system. Engineeringin Agriculture, Environment and Food, 11(4):202 – 210, October 2018.doi: 10.1016/j.eaef.2018.05.002

[64] A. S. A. Rabby, S. Haque, S. Islam, S. Abujar, and S. A. Hossain.Bornonet: Bangla handwritten characters recognition using convolu-tional neural network. Procedia Computer Science, 143:528 – 535,2018. doi: 10.1016/j.procs.2018.10.426

[65] M. A. Rahaman, M. Jasim, M. H. Ali, and M. Hasanuzzaman. Real-time computer vision-based bengali sign language recognition. In2014 17th International Conference on Computer and InformationTechnology (ICCIT), pp. 192–197. IEEE, Dec 2014. doi: 10.1109/ICCITechn.2014.7073150

[66] J. L. Raheja, S. Deora, and A. Chaudhary. Cross border intruderdetection in hilly terrain in dark environment. Optik, 127(2):535 – 538,January 2016. doi: 10.1016/j.ijleo.2015.08.234

[67] H. G. Rai, K. Jonna, and P. R. Krishna. Video analytics solution fortracking customer locations in retail shopping malls. In Proc. of the17th ACM SIGKDD International Conference on Knowledge Discoveryand Data Mining, KDD ’11, pp. 773–776. ACM, NY, USA, 2011. doi:10.1145/2020408.2020533

[68] H. T. Rauf, B. A. Saleem, M. I. U. Lali, M. A. Khan, M. Sharif, andS. A. C. Bukhari. A citrus fruits and leaves dataset for detection andclassification of citrus diseases through machine learning. Data inBrief, 26:104340, October 2019. doi: 10.1016/j.dib.2019.104340

[69] Z. Ren, Y. Jae Lee, and M. S. Ryoo. Learning to anonymize facesfor privacy preserving action detection. In The European Conferenceon Computer Vision (ECCV). Springer, 2018. doi: 10.1007/978-3-030-01246-5 38

[70] D. Roy and K. M. C. Snatch theft detection in unconstrained surveil-lance videos using action attribute modelling. Pattern RecognitionLetters, 108:56 – 61, June 2018. doi: 10.1016/j.patrec.2018.03.004

[71] M. S. Ryoo, B. Rothrock, C. Fleming, and H. J. Yang. Privacy-preserving human activity recognition from extreme low resolution. In

AAAI Conference on Artificial Intelligence. AAAI Press, 2017.[72] S. Z. H. Shah, H. T. Rauf, M. IkramUllah, M. S. Khalid, M. Farooq,

M. Fatima, and S. A. C. Bukhari. Fish-pak: Fish species datasetfrom pakistan for visual features based classification. Data in Brief,27:104565, December 2019. doi: 10.1016/j.dib.2019.104565

[73] R. Singh and S. J. Jackson. From margins to seams: Imbrication,inclusion, and torque in the aadhaar identification project. In Proc. ofthe 2017 CHI Conference on Human Factors in Computing Systems,CHI ’17, p. 4776–4824. ACM, NY, USA, 2017. doi: 10.1145/3025453.3025910

[74] M. Skirpan and T. Yeh. Designing a moral compass for the future ofcomputer vision using speculative analysis. In IEEE Conference onComputer Vision and Pattern Recognition Workshops (CVPRW), pp.1368–1377, July 2017. doi: 10.1109/CVPRW.2017.179

[75] L. T. Smith. Decolonizing methodologies: Research and indigenouspeoples. Zed Books Ltd., 2013.

[76] P. Speciale, J. L. Schonberger, S. B. Kang, S. N. Sinha, and M. Polle-feys. Privacy preserving image-based localization. In The IEEE Confer-ence on Computer Vision and Pattern Recognition (CVPR), pp. 5488–5498. IEEE, June 2019. doi: 10.1109/CVPR.2019.00564

[77] U. N. Statistics Division. Standard country or area codes for statisticaluse (M49), 2017.

[78] S. Suwajanakorn, S. M. Seitz, and I. Kemelmacher-Shlizerman. Syn-thesizing obama: Learning lip sync from audio. ACM Trans. Graph.,36(4):95:1–95:13, July 2017. doi: 10.1145/3072959.3073640

[79] G. Varma, A. Subramanian, A. Namboodiri, M. Chandraker, andC. Jawahar. IDD: A dataset for exploring problems of autonomousnavigation in unconstrained environments. In 2019 IEEE Winter Con-ference on Applications of Computer Vision (WACV), pp. 1743–1751.IEEE, 2019. doi: 10.1109/WACV.2019.00190

[80] Y. Wang and M. Kosinski. Deep neural networks are more accuratethan humans at detecting sexual orientation from facial images. Journalof personality and social psychology, 114(2):246–257, February 2018.doi: 10.1037/pspa0000098

[81] E. Zakharov, A. Shysheya, E. Burkov, and V. Lempitsky. Few-shotadversarial learning of realistic neural talking head models. arXiv,arXiv:1905.08233, 2019.

[82] R. Zemel, Y. Wu, K. Swersky, T. Pitassi, and C. Dwork. Learning fairrepresentations. In International Conference on Machine Learning,vol. 28 of Proceedings of Machine Learning Research, pp. 325–333.PMLR, Atlanta, Georgia, USA, 2013. doi: 10.5555/3042817.3042973

[83] J. Zou and L. Schiebinger. AI can be sexist and racist — it’s time tomake it fair. Nature, 559(7714):324–326, July 2018. doi: 10.1038/d41586-018-05707-8